Title: IMWM: Integrating Intuition with World Models for Enhanced Latent Planning

Abstract

While leveraging learned latent world models offers a viable pathway for controlling agents directly from raw pixels, relying on world models in isolation proves insufficient. Our experimental findings demonstrate this limitation clearly: even when equipped with a flawless world model—simulated by substituting the learned forward predictor with an idealized rollout of the actual environment dynamics—a sample-based planner operating under finite budget constraints still struggles on certain tasks. This suggests that the primary bottleneck may reside within the search process itself rather than in the precision of the world model.

To bridge this gap, we introduce IMWM (Intuition Model + World Model), a framework that augments the world model with an intuition model. The latter is trained on demonstration data to identify high-potential actions. These two components interact via three streamlined mechanisms: (i) Retrieval Initialization, which seeds the planner’s action proposals using retrieved demonstrations; (ii) Hybrid Cost, which merges the intuition model’s score with the cost derived from world-model rollouts; and (iii) a Reliability Gate, which dynamically calibrates the planner’s reliance on intuition based on the specific context.

We evaluated IMWM across four pixel-based goal-reaching environments: Two-Room, Reacher, Push-T, and OGBench-Cube. The results show that IMWM outperforms planners relying solely on world models in terms of mean success rates across all four tasks. The most significant improvements were observed in the Two-Room environment, where success rates reached 99.2% (an increase of 11.5 percentage points), and in OGBench-Cube, where success rates hit 94.7% (a substantial gain of 28.5 percentage points).

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC